Async serving: ROUTER socket, predict thread and latest-wins request queue

[khanhnd61-vr]

Summary

vla-server was a single-threaded REQ/REP loop: one request, one reply, and the socket was deaf while the model ran. A robot client therefore paused for a full round trip per chunk. This PR makes the server asynchronous without changing anything for existing clients.

• ROUTER socket + predict thread. The socket thread receives, validates and decodes; the worker is the only caller of vla::predict. Replies travel back over an inproc PUSH/PULL and carry the peer's routing envelope. A REQ client sees exactly the exchange it did before; a DEALER client can keep several requests in flight.
• --queue latest|fifo (src/serving/request_queue.h, header-only). latest (default) keeps one pending request per client and answers a superseded one at once with error="superseded", so the model always works on each robot's freshest observation and several robots share a server fairly. fifo serves every request in order for benchmarks.
• latency_ms_queue added to PredictResponse: how long a request waited for the predict thread. Malformed requests are rejected on the socket thread, so errors come back within milliseconds mid-predict.
• Docs (USAGE.md, ARCHITECTURE.md, README rollout section), CHANGELOG entry, and tests/test_request_queue.cpp.

Companion PRs: the lerobot fork gains lerobot-vla-cpp --mode=async (khanhnd61-vr/lerobot), and vla.simd gets its session lock out of the predict path (cair-vinuni/vla.simd).

Test plan

• test_request_queue (ctest): coalescing per client, no cross-client coalescing, FIFO order and depth cap, drain-then-stop, blocking pop.
• End to end on an RTX 3060 with the SmolVLA LIBERO checkpoint: legacy REQ client unchanged; a 5-request DEALER burst serves the first and the newest (queue wait = one predict) and answers the three stale ones in 26 ms; two DEALER clients both served; a malformed request answered in 6 ms during a 147 ms predict; REQ and DEALER interleaved; --queue fifo serves all five in order with queue waits growing by one predict each; SIGINT shuts down cleanly with the served/superseded summary.
• lerobot-vla-cpp async client against this server at 30 fps for 8 s: 229 of 239 ticks executed an action, the 10 starved ticks were the initial fill, 0.05 ms median control-thread work per tick.